A Bayesian approach for subgroup analysis.

Area of Science:

• Statistical modeling
• Biostatistics
• Machine learning

Background:

• Existing penalization methods for subgroup analysis in regression models with subject-specific intercepts rely on pairwise intercept comparisons.
• These methods often employ concave penalty functions, mirroring variable selection techniques.
• Bayesian approaches are prevalent in variable selection, suggesting their potential utility in subgroup analysis.

Purpose of the Study:

• To develop a Bayesian hierarchical model for identifying homogeneous subgroups based on subject-specific intercepts in regression analysis.
• To automatically detect and group subjects with similar intercept values.
• To offer a computationally efficient alternative to existing penalization methods.

Main Methods:

• A Bayesian hierarchical model is proposed, incorporating prior structures for pairwise intercept differences.
• A Gibbs sampling algorithm is utilized for hyperparameter selection and simultaneous estimation of intercepts and covariate coefficients.
• The model is evaluated using simulation studies and applied to the Cleveland Heart Disease Dataset.

Main Results:

• The proposed Bayesian method effectively identifies homogeneous subgroups.
• The Gibbs sampling algorithm provides computationally efficient parameter estimation and hyperparameter selection.
• The method demonstrates superior performance compared to penalization techniques, especially for large datasets.

Conclusions:

• The developed Bayesian hierarchical model offers an effective and computationally efficient approach for subgroup analysis.
• This method automatically detects homogeneous subgroups by analyzing pairwise intercept differences.
• The findings suggest a promising Bayesian framework for subgroup identification in regression modeling.